FLPS model

  • bu11: -1; bu12: 0.5; by1: 1; by2: 0.5; b0 around 0.4; b11 around -0.2; a11 around 0.2

  • 4 latent models (Rasch, 2PL, GPCM, GRM)

  • Conditions

    • sample size: 500, 1000
    • N itmes: 50 100 200

  • 2 chains with 5000 iterations (2000 warmup)

  • Outcome: bias (Estimate - True value). Zero = no bias.

Results

structural model

  • bu11 always overestimated; bu12 always underestimated.
  • factor loading (discrimination) highly overestimated (due to priors? Normal(1, 1) with int<lower=0> lambda )
  • eta (individual latent scores) underestimated

measurement model

correlation plot

  • As expected, some parameters are correlated.

Rasch

2PL

GRM

GPCM

with lognormal prior

  • Here, I only used the measurement model with covariates to see if the different priors might address overestimated factor loadings.

  • Using lognormal priors for factor loadings, the estimates seem to be more stable around 0.